Visual Reinforcement Learning for Dynamic Object Detection

Abstract. Object detection is a widely studied task in computer vision. Current methods often focus on images captured from appropriate viewpoints. However, there is a large disparity between objects observed from different viewpoints in the real world. Dynamic Object Detection (DOD) method automatically adjusts the camera viewpoint in a visual scene to sequentially find optimal viewpoints. Currently, the DOD tasks are usually modeled as a sequential decision-making problem and solved using reinforcement learning methods. Existing approaches face challenges with sparse rewards and training instability. To tackle these issues, we proposed a single-step reward function and a lightweight network, respectively. The single-step reward function, which provides timely feedback, gives an efficient training process for DOD tasks. The lightweight network with few parameters can ensure the stability of the training process. To evaluate the effectiveness of our method, we developed a simulation dataset based on UE4, which consists of 1800 training images and 450 testing images. The dataset includes five object categories: vans, cars, trailers, box trucks and SUVs. Experiments demonstrate that our method outperforms SOTA object detectors on our simulation dataset. Specifically, the average precisions(APs) are improved from 89.1% to 96.0% when using the YOLOv8 object detector.